rsrl 0.8.1

A fast, extensible reinforcement learning framework in Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168

use crate::{
    fa::{transforms, StateActionUpdate, StateUpdate, GradientUpdate, ScaledGradientUpdate},
    params::{Buffer, BufferMut},
    Differentiable,
    Function,
    Handler,
    Parameterised,
};

/// Composition of an FA and a differentiable transform.
#[derive(Clone, Debug, Parameterised)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct Composition<F, T> {
    #[weights]
    pub fa: F,
    pub transform: T,
}

impl<F, T> Composition<F, T> {
    pub fn new(fa: F, transform: T) -> Self { Composition { fa, transform } }
}

///////////////////////////////////////////////////////////////////////////////////////////////////
// Common
///////////////////////////////////////////////////////////////////////////////////////////////////
impl<Args, F, T> Function<Args> for Composition<F, T>
where
    F: Function<Args>,
    T: transforms::Transform<F::Output>,
{
    type Output = T::Output;

    fn evaluate(&self, args: Args) -> T::Output {
        self.transform.transform(self.fa.evaluate(args))
    }
}

impl<F, T, J> Handler<GradientUpdate<J>> for Composition<F, T>
where
    F: Handler<GradientUpdate<J>>,
    J: Buffer,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(&mut self, msg: GradientUpdate<J>) -> Result<Self::Response, Self::Error> {
        self.fa.handle(msg)
    }
}

impl<F, T, J> Handler<ScaledGradientUpdate<J>> for Composition<F, T>
where
    F: Handler<ScaledGradientUpdate<J>>,
    J: Buffer,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(&mut self, msg: ScaledGradientUpdate<J>) -> Result<Self::Response, Self::Error> {
        self.fa.handle(msg)
    }
}

///////////////////////////////////////////////////////////////////////////////////////////////////
// Implement V(s)
///////////////////////////////////////////////////////////////////////////////////////////////////
// TODO: implement for any fa/transform combination, not just those with output
// f64.
impl<'s, S, F, T> Differentiable<(&'s S,)> for Composition<F, T>
where
    F: Differentiable<(&'s S,), Output = f64>,
    T: transforms::Transform<F::Output, Output = f64>,

    F::Output: Clone,
{
    type Jacobian = F::Jacobian;

    fn grad(&self, args: (&'s S,)) -> Self::Jacobian {
        let gx = self.fa.evaluate(args);
        let f_gx = self.transform.grad(gx);

        self.fa.grad(args).map_into(|x| f_gx * x)
    }

    fn grad_log(&self, args: (&'s S,)) -> Self::Jacobian {
        let gx = self.fa.evaluate(args);
        let fgx = self.transform.transform(gx.clone());
        let f_gx = self.transform.grad(gx);

        self.fa.grad(args).map_into(|x| f_gx * x / fgx)
    }
}

impl<'s, S, F, T> Handler<StateUpdate<&'s S, f64>> for Composition<F, T>
where
    F: Differentiable<(&'s S,), Output = f64> + Handler<StateUpdate<&'s S, f64>>,
    T: transforms::Transform<f64, Output = f64>,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(&mut self, msg: StateUpdate<&'s S, f64>) -> Result<Self::Response, Self::Error> {
        let gx = self.fa.evaluate((msg.state,));
        let f_gx = self.transform.grad(gx);

        self.fa.handle(StateUpdate {
            state: msg.state,
            error: msg.error * f_gx,
        })
    }
}

///////////////////////////////////////////////////////////////////////////////////////////////////
// Implement Q(s, a)
///////////////////////////////////////////////////////////////////////////////////////////////////
impl<'s, 'a, S, A, F, T> Differentiable<(&'s S, &'a A)> for Composition<F, T>
where
    F: Differentiable<(&'s S, &'a A), Output = f64>,
    T: transforms::Transform<F::Output, Output = f64>,

    F::Output: Clone,
{
    type Jacobian = F::Jacobian;

    fn grad(&self, args: (&'s S, &'a A)) -> Self::Jacobian {
        let gx = self.fa.evaluate(args);
        let f_gx = self.transform.grad(gx);

        self.fa.grad(args).map_into(|x| f_gx * x)
    }

    fn grad_log(&self, args: (&'s S, &'a A)) -> Self::Jacobian {
        let gx = self.fa.evaluate(args);
        let fgx = self.transform.transform(gx.clone());
        let f_gx = self.transform.grad(gx);

        self.fa.grad(args).map_into(|x| f_gx * x / fgx)
    }
}

impl<'s, S, A, F, T> Handler<StateActionUpdate<&'s S, A, f64>> for Composition<F, T>
where
    A: std::borrow::Borrow<f64>,
    F: Differentiable<(&'s S, f64), Output = f64> + Handler<StateActionUpdate<&'s S, A, f64>>,
    T: transforms::Transform<f64, Output = f64>,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(
        &mut self,
        msg: StateActionUpdate<&'s S, A, f64>,
    ) -> Result<Self::Response, Self::Error>
    {
        let gx = self.fa.evaluate((msg.state, *msg.action.borrow()));
        let f_gx = self.transform.grad(gx);

        self.fa.handle(StateActionUpdate {
            state: msg.state,
            action: msg.action,
            error: msg.error * f_gx,
        })
    }
}